OpenAI has upgraded its Responses API, signaling a shift in the development of AI agents, while new memory architectures are emerging to cut costs and improve performance, according to recent reports from VentureBeat. The updates to the API, which allows developers to access multiple agentic tools with a single call, include Server-side Compaction and Hosted Shell. Simultaneously, the industry is exploring alternative memory architectures, such as "observational memory," to enhance the capabilities of AI agents.
The advancements in AI agent technology come as the industry grapples with the challenges of scaling these systems. According to VentureBeat Contributing Editor Taryn Plumb, "what has been missing are the fundamentals required to scale." The limitations of existing systems, particularly in long-running, tool-heavy agents, have prompted the search for more efficient solutions. "RAG isn't always fast enough or intelligent enough for modern agentic AI workflows," noted Sean Michael Kerner of VentureBeat.
One promising development is "observational memory," an open-source technology developed by Mastra, which prioritizes persistence and stability over dynamic retrieval. This approach aims to reduce costs and improve performance, potentially cutting AI agent costs by a factor of 10 and outperforming RAG on long-context benchmarks, according to VentureBeat.
The application of AI in fraud detection also highlights the rapid advancements in the field. Mastercard's Decision Intelligence Pro (DI Pro) uses sophisticated AI models to analyze individual transactions and identify suspicious activity in milliseconds, according to VentureBeat. Johan Gerber, Mastercard's EVP, explained that DI Pro focuses on "each transaction and the risk associated with it." This technology is crucial, as Mastercard processes approximately 160 billion transactions annually, with peak periods experiencing up to 70,000 transactions per second.
The evolution of AI agents is also being explored in the context of Global Business Services (GBS). While the potential for agentic AI to transform enterprises is significant, the actual deployment has lagged behind the hype. As N. Shashidar of EdgeVerve noted in VentureBeat, "As with many new technologies, rhetoric has outpaced deployment in this case."
The MIT Technology Review has also launched a new weekly newsletter, "Making AI Work," to explore the practical applications of generative AI across various sectors. The newsletter will examine specific use cases and provide insights into how professionals can apply AI in their everyday work.
Discussion
AI Experts & Community
Be the first to comment